Deformations in isotropic, strain-free polymer glasses are usually completely recoverable (at the test temperature or after warming to T-g), in sharp contrast with the behavior of low molecular weight glasses and crystals. The apparent 'plastic strain' which remains at the end of a creep or stress relaxation experiment does not recover at the test temperature, but only after the sample is heated. It is proposed that the long time scales needed for entanglement reorganization in the glass are responsible for this delayed recovery. A phenomenological network model for thermally activated strain recovery in polymer glasses is analyzed. A superposition relation between the stress and the strain history using a KWW (stretched exponential) memory kernel is employed. The recovery of plastic (i.e. residual) strain in non-crosslinked amorphous thermoplastics is a two-step process that may be interpreted in terms of the network model. In particular, recovery at sub-T-g temperatures is associated with entanglement slippage, while recovery near-T-g is believed to involve reorganization at or near chain ends.